跳转至

文章背景与核心概要

本文介绍了由 Yan Zhou 等研究人员提出的全新测试时推理（Test-time reasoning）优化框架——CoBa（Compute-Balanced Routing）。在现有的大语言模型应用中，测试时扩展通常面临单一维度的计算资源竞争（如盲目增加采样数量、延长思维链或调用昂贵的评估器）。为了突破这一效率瓶颈，CoBa 将测试时推理建模为一个动态的计算分配优化问题，能够智能决定下一个计算单元应当投入到生成、验证还是终止中。

该策略的核心技术在于三步流水线：首先生成一小组合格候选；接着对所有候选应用低成本的泛化验证；最后将高价值或存在不确定性的候选路由至更强健的深层验证流程。实验涵盖了 MATH-500、AIME 2024/2025、AMC 2023 及过程符号推理共 3,129 个样本评估，结果表明 CoBa-Routed-Strong 达到了 85.13% 的宏平均准确率，在保持与自我评估加权投票代理（85.20%）几乎一致性能的同时，大幅减少了 49.1% 的参数加权 Token 消耗，为高效的本地推理系统开辟了极具前景的技术路径。

CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing

Authors: Yan Zhou, Yue Ouyang, Kaiyang Zheng, Suncheng Xiang
Submitted: August 7, 2026
Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.07424 [cs.AI]
DOI: 10.48550/arXiv.2608.07424

CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing

Authors: Yan Zhou, Yue Ouyang, Kaiyang Zheng, Suncheng Xiang
Submitted: August 7, 2026
Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.07424 [cs.AI]
DOI: 10.48550/arXiv.2608.07424

Executive Summary

测试时扩展（Test-time scaling）通常需要将更多计算资源分配给单一维度，例如采样更多的解决方案、扩展思维链，或者部署更强的评估器。由于这些策略在固定的推理预算下相互竞争，CoBa（计算均衡路由，Compute-Balanced Routing） 将测试时推理构想为一个最优计算分配挑战——即动态决定下一个计算单元是用在生成、验证还是终止上。

CoBa 实现了一种计算均衡的路由策略，其步骤如下： 1. 生成初始的一小组合候选集。 2. 对这些候选集广泛应用低成本验证。 3. 将不确定或高价值的候选路由至更稳健、更强大的验证流程中。

在跨越 MATH-500、AIME 2024/2025、AMC 2023 以及过程符号推理的 3,129 个示例生成器评估 中： * CoBa-Routed-Strong 实现了 85.13% 的宏平均准确率，紧密逼近自我评估加权投票代理（85.20%），同时节省了 49.1% 的参数加权 Token。 * 它在宏平均准确率上与 16 取最优（best-of-16）多数投票相差不到 0.01 个百分点，同时消耗了 58.9% 的参数加权 Token。

成对自助法（bootstrap）测试证实，相比标准的单样本解码，该方法取得了显著的性能提升，为高效的本地推理系统指明了一条充满希望的发展路线。

Executive Summary

Test-time scaling typically requires allocating more compute to a single axis, such as sampling a greater number of solutions, extending chains of thought, or deploying a stronger evaluator. Because these strategies compete under a fixed inference budget, CoBa (Compute-Balanced Routing) frames test-time reasoning as an optimal compute-allocation challenge—determining dynamically whether the next unit of compute is best spent on generation, verification, or stopping.

CoBa implements a compute-balanced routing policy that: 1. Generates an initial small set of candidates. 2. Applies low-cost verification broadly across them. 3. Routes uncertain or high-value candidates to more robust, stronger verification processes.

Across 3,129 example-generator evaluations spanning MATH-500, AIME 2024/2025, AMC 2023, and procedural symbolic reasoning: * CoBa-Routed-Strong achieves an 85.13% macro accuracy, closely matching a self-evaluation weighted-voting proxy (85.20%) while utilizing 49.1% fewer parameter-weighted tokens. * It matches best-of-16 majority voting within 0.01 macro-accuracy points while consuming 58.9% fewer parameter-weighted tokens.

Paired bootstrap tests confirm significant performance gains over standard single-sample decoding, highlighting a promising roadmap for efficient local reasoning systems.

Abstract

测试时扩展通常通过在单一维度上消耗更多计算来实现：采样更多解决方案、扩展思维链，或者应用更强的评估器。在固定的推理预算下，这些选择会相互竞争。本文将测试时推理建模为一个计算分配问题，系统必须在此问题中决定下一个计算单元是应用于生成、验证还是终止。我们推出了 CoBa，这是一种计算均衡的路由政策，它首先获取一小组合候选，广泛应用低成本验证，并将不确定或高价值的候选路由至更强的验证中。在涵盖 MATH-500、AIME 2024/2025、AMC 2023 和过程符号推理的 3,129 个示例生成器评估中，CoBa-Routed-Strong 达到了 85.13% 的宏平均准确率，在统计上与 85.20% 的自我评估加权投票代理相匹配，同时使用了少 49.1% 的参数加权 Token。它还与 16 取最优多数投票在 0.01 的宏平均准确率差距内相匹配，同时使用了少 58.9% 的参数加权 Token；成对测试在大幅增加成本的情况下保留了微弱的 16 取最优优势。成对自助法测试显示，相比单样本解码有显着增益，而与候选池最优解（pool oracle）之间剩余的差距则暴露出实现更精准路由的空间。对于本地推理系统而言，测试时扩展变成了一个关于“在哪里投入下一个计算最有价值”的问题。

Abstract

Test-time scaling is often implemented by spending more compute along one axis: sampling more solutions, extending a chain of thought, or applying a stronger evaluator. Under a fixed inference budget, these choices compete. This paper formulates test-time reasoning as a compute-allocation problem in which a system must decide whether the next unit of compute should be spent on generation, verification, or stopping. We introduce CoBa, a compute-balanced routing policy that first obtains a small set of candidates, applies cheap verification broadly, and routes uncertain or high-value candidates to stronger verification. On 3,129 example-generator evaluations spanning MATH-500, AIME 2024/2025, AMC 2023, and procedural symbolic reasoning, CoBa-Routed-Strong reaches 85.13% macro accuracy, statistically matching a self-evaluation weighted-voting proxy at 85.20% while using 49.1% fewer parameter-weighted tokens. It also matches best-of-16 majority voting within 0.01 macro-accuracy points while using 58.9% fewer parameter-weighted tokens; paired tests retain a small best-of-16 edge at substantially higher cost. Paired bootstrap tests show significant gains over single-sample decoding, while the remaining gap to the pool oracle exposes headroom for sharper routing. For local reasoning systems, test-time scaling becomes a question of where the next computation is most valuable.

Access Full-Text & Resources

Access Full-Text & Resources

References & External Tools

References & External Tools